perf: compact 3D CPML state

[Vcholerae1]

Summary

• keep compact CPML shape/pack/unpack bookkeeping as an internal grid utility
• keep the public EM3D state representation full-grid and backend-independent
• migrate non-Born CUDA propagation and adjoint scratch to axis-compact CPML storage internally
• scope compact PML indexes to active branches and specialize adjoint raw accessors by compile-time axis/sign
• preserve full-grid callback views, CPU/reference continuation, Born output compatibility, Debye behavior, and public API
• leave TM2D full-grid after its multi-shot compact prototype regressed
• remove stale quality-workflow references to the benchmark script deleted on main

Results

• internal CPML working-state storage: 52,800 -> 25,600 bytes on the parity workload (-51.52%)
• Nsight Compute E adjoint kernel: 64.22 -> 46.66 us (-27.3%)
• Nsight Compute H adjoint kernel: 67.23 -> 49.50 us (-26.4%)
• 65^3, 1200-step backward CUDA median improved 16.95% versus the pre-specialization compact implementation
• public returned states remain full-grid; memory claims exclude returned-state packaging and callback expansion

Verification

• Ruff format/check: pass (83 files)
• CPU non-numerical: 288 passed, 2 skipped
• CPU numerical: 38 passed
• CUDA numerical: 48 passed
• targeted continuation/Debye/Born/callback/batched-float64 guards: 10 passed

TM2D and Born computational layouts are intentionally unchanged.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.